Dental care

An AI receptionist for a dental clinic: what it should answer – and what it never should

A dental front desk runs one phone queue with six completely different calls in it. Only some of them should reach a voice agent – and one of them it should never assess on its own.

A woman in a dark sweatshirt stands alone at the end of a kitchen table in an apartment at blue hour before sunrise, a mobile phone held to her right ear and her other hand flat on the table top, cold light falling in from the window on the left and a warm strip of under-cabinet lighting along the counter on the right

Monday, 08:00

The phone is ringing. A patient is standing at the counter paying for the treatment she just had. The receptionist has the handset in one hand and the card terminal in the other, and in surgery 2 the dentist needs a chart nobody has pulled yet.

At the same time, someone who woke at four with a molar throbbing is calling for the first time. She gets a busy tone. She calls the next clinic on the search results instead, and that call does not come back.

This is not a problem of nobody answering the phone. It is a problem of one queue having to carry calls of wildly different value, urgency and risk – and of one person at the front desk having to sort them in real time, between two patients who already have her attention.

So the question isn't whether an AI should answer your phone. The question is which of those calls it should handle, which it should take down and pass on, and which it should not go anywhere near.

Six calls in one queue

Call Volume Risk Who should take it
New patient Low Low The agent, end to end
Rescheduling and cancellations High Low The agent, end to end
Recall confirmations High Low The agent, end to end
Practical questions Medium Low The agent, end to end
Billing, insurance and price Medium Medium The agent captures, a human answers
Acute pain, swelling, trauma Low High Triage, then a human or emergency dental care

Notice how little the six have in common. One needs clinical judgement. One needs access to the accounts. Two are pure calendar logic. And they arrive in random order on the same number, where the one person who has to sort them already has a patient in front of her.

What the agent should own outright

Booking, rescheduling and cancelling is pure calendar logic. An agent that can see open slots and write a new one handles the call to completion – no callback, no note left at the desk. The same goes for confirming and moving recalls, one of the most common reasons the phone rings at all, and for practical questions: opening hours, address, parking, what to bring, whether the caller is registered with you.

New patients should be closed out, not merely captured. The difference between "we'll call you back tomorrow" and "you're booked for Thursday at 09:20" is the difference between a lead and a patient. An agent that answers at 21:00 on a Sunday and books there and then picks up calls the clinic effectively had no access to before – and the value per answered call is higher at a dental practice than in most places, because a registered patient is a patient for years.

On the practical side: a Threll books directly against cal.com today. A hookup to the practice management or diary system your clinic already runs is something we set up per clinic – tell us which system you're on and we'll take it from there.

Billing and insurance, by contrast, should be captured rather than answered. The agent should record what the question concerns, which invoice or treatment it relates to, and who needs calling back. It should not guess at coverage or patient contributions. A wrong answer here costs more trust than it saves time.

The rule of thumb: the agent should own the calls where the right answer lives in a calendar or in the clinic's own routines. Not the calls where the right answer takes clinical or financial judgement.

Want to talk it through with someone who has done this before?

Fifteen minutes on a call. We will tell you honestly whether a voice agent fits your setup – and what has to be settled before it goes live.

15 minutes · no obligation · pick your own time

Acute pain is triage, not booking

This is the section that decides whether the whole thing is defensible.

An agent can safely collect: how long it has been going on, whether there is swelling in the face or towards the throat, whether the patient has a fever, whether swallowing or breathing hurts, whether a tooth has been knocked out or loosened in an injury, whether the pain is keeping the patient awake.

It must not assess. It should have three outcomes and only three: a same-day slot, a transfer to a human now, or a referral to out-of-hours emergency dental care when the clinic is closed. Which of the three is triggered by what, who owns the case afterwards, and what happens when nobody picks up, is a design exercise in its own right.

Three hard rules belong in the configuration:

  1. Never clinical advice. No medication, no self-treatment, no view on what the pain is caused by.
  2. Never reassurance. "That doesn't sound too serious" is a sentence the agent should not be able to produce.
  3. Never the calendar on a red flag. An avulsed permanent tooth, swelling towards the eye or throat, difficulty breathing, high fever: a human or an emergency service immediately. An avulsed permanent tooth is time-critical in minutes, not hours – that call should never end up in a booking dialogue.

It is worth testing this part harder than everything else before a clinic goes live. The rest of the system fails expensively. This part fails dangerously.

The cancelled hour

The under-sold half of this has nothing to do with answering the phone at all.

An appointment that frees up at 14:00 tomorrow is lost revenue unless someone fills it. The waiting list exists – often in a spreadsheet, or in the head of whoever has worked there longest – but nobody at the desk has time to ring through it between two patients. So the slot stays empty.

An agent doesn't mind that there are twelve names on the list. It calls or texts in priority order until someone says yes, and books them in on the spot. The same mechanism covers the other outbound jobs nobody gets to: chasing unconfirmed recalls, and patients who haven't been in for two years.

It is also the lowest-risk part. No patient has to discuss symptoms with a machine on an outbound call about a free slot tomorrow. If you want to try this cautiously, start here.

The phone hears health data

"I've got a throbbing molar, upper left, and it's kept me awake for three nights" is health information. Under GDPR it is a special category of personal data, and the moment a vendor answers the phone on the clinic's behalf, that vendor is processing it.

There is a real difference between automating a dental clinic's phone and a hair salon's, and this is where it sits. Questions worth putting to any vendor, us included:

  • Is there a data processing agreement, and who are the sub-processors? The model provider, the telephony provider and any transcription service are all in the chain.
  • Where is audio and transcript data stored, and for how long?
  • Is audio or transcript data used to train models?
  • How is the patient told that the call is being handled by an agent, and whether it is recorded?
  • How does the setup sit inside the information-security and record-keeping obligations your clinic is already bound by, whoever answers the phone?

If the answer to all of that is "we're GDPR compliant" and nothing further, it isn't an answer. These are questions with concrete, checkable answers, and a clinic should have them in writing before anyone connects anything to the phone line.

What "it works" looks like after 30 days

Not a good feeling and not a testimonial. Five numbers the clinic can pull itself:

  • Share of calls answered, split by inside and outside opening hours. Outside hours is where the big change sits.
  • Share of calls closed without a transfer. How much actually left the queue, rather than just moving along it.
  • New patients booked from calls that would previously have gone unanswered.
  • Slots filled from the waiting list, measured in hours per week.
  • Number of acute calls routed to a human. This number should not be zero. If it is, the threshold is set wrong, and it's worth finding out why before somebody is harmed by it.

If the first four haven't moved after a month, the problem isn't that the technology is immature. It's that it has been set up to solve something other than what the clinic actually struggles with.

Where to start

Start where the risk is lowest and the effect is quickest to measure: the phone outside opening hours, and the waiting list. Booking against cal.com works today, and the acute-call routing should be set up alongside someone at the clinic who takes those calls now.